FMA torso anatomy: filled smooth triangle-mesh cockpit#57
Merged
Conversation
The Railway cockpit-server build now compiles q2-ndarray with a v4 floor (CARGO_BUILD_RUSTFLAGS=-C target-cpu=x86-64-v4), flipping on target_feature="avx512f" so ndarray's simd.rs dispatch selects the native simd_avx512 backend. BF16 + AMX tile GEMM ride ndarray's runtime autodetect polyfill (simd_caps + the AMX arch_prctl XTILEDATA enable + CPU-model detect) — lit only when the host has them, AVX2/ scalar fallback always compiled. Documents the requirement (a v4 floor makes the binary require AVX-512 at run time; SIGILL otherwise) and the x86-64-v3 fallback for deploy targets that may lack it. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TzqvDqbFRzyx17EkLKBoZF
Pivot the BodyParts3D bake from the part-of trunk subtree to the is_a TYPE decomposition. part-of is regional (walking up a muscle hits chest-wall -> thorax, never a muscular system) and carries non-canonical names; is_a classifies every structure to its canonical type (muscle organ / artery / vein / nerve / bone organ / cartilage / ...), ships canonical names, and is a mesh superset of part-of (2234 vs 1258 FJ, +976) with finer organ segmentation (no single aorta/heart concept; split into ascending/arch/descending/abdominal, each its own mesh). - bake_torso_splat.py v4, is_a-primary: tissue resolved by walking the is_a type tree (O(1), cached); conventional atlas palette (bone tan, muscle red, artery red, vein blue, nerve yellow, viscera distinct); depth-peel opacity (wall/muscle translucent, deep organs/vessels/bone solid) so inner structures show through. - Addressing: each structure's is_a DistinguishedName path (upper- ontology stripped) materialises to a numeric container:identity GUID (container = tissue class). Stored per node alongside centroid+bbox. - Whole body, no spatial clip — region focus (torso, an organ) is a future select -> camera-zoom on the SoA centroid/bbox, not a bake filter. - Asset: 602,341 gaussians / 1658 is_a structures (414 arteries, 382 muscles, 221 veins, 203 bones, 126 nerves, full viscera), 12.6 MB SPL2 (wire unchanged; existing decoders parse it as-is). - Source: BodyParts3D 4.0 (DBCLS), CC-BY 4.0 / CC-BY-SA 2.1 JP — permissive, commercial-safe. Gap-analysis research note + plan update. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TzqvDqbFRzyx17EkLKBoZF
Re-render the /torso CPU EWA turntable (20 frames) from the v4 is_a-primary whole-body asset: a standing figure, tissue-coloured (muscle red, veins blue down the limbs), depth-peel translucency revealing deeper structure, lit from the baked surface normals. Driver orientation fixed (+90 about X; the body had been landing head-down). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TzqvDqbFRzyx17EkLKBoZF
/torso-live (TorsoSplat) and /torso-map (TorsoMap) now decode the SPL2 normal + opacity and render through a lit ShaderMaterial — hemisphere (sky/ground) ambient + key diffuse + soft fill from the surface normal, with per-gaussian depth-peel opacity. Same recipe as the CPU driver, so all three views agree; orientation matches the +90 flip (no more points.rotation hack). Performance work for the 602K-point whole-body load: - LazyLock: geometry built ONCE; selection + throttle drive uniforms / drawRange, never a rebuild or re-upload. A shuffled index makes an adaptive drawRange prefix a uniform spatial subsample (thins evenly, not by structure order). - Adaptive-FPS: an EMA of the frame time thins the draw range and drops pixelRatio before a stutter lands, recovering when frames are cheap. Active fraction + fps surfaced in-view (never silent decimation). cockpit build (tsc + vite) green. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TzqvDqbFRzyx17EkLKBoZF
Per-triangle oriented surfels (one disk per face, not per vertex) sized by per-structure k-NN local spacing (~0.55x the 3-nearest distance) -- the metric Sigma-fit, replacing the hand-rolled sqrt(stride) bridge that over-inflated disks into marbles. SPL3 wire (22 B body adds a per-surfel scale byte; header f32 becomes scale_max, the disk-radius dequant ref). export_isa_classification.py emits the full FJ->tissue/system/GUID/DN table (1658 concepts / 2234 meshes) for the certified- gaussian session, reusing the same is_a walk so both renders share one classifier. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TzqvDqbFRzyx17EkLKBoZF
…fog) The "ghost" was the renderer, not the asset: the soft-Gaussian point material composited alpha tails through the translucent depth-peel surfels -> halo + fog. All three views (/torso, /torso-live, /torso-map) now render OPAQUE depth-tested surfels, sized per-gaussian from the SPL3 scale byte: nearest surface wins, hard edge, no alpha accumulation -> crisp. uFloor gates the skin/flesh shell out (press S on /torso-live to toggle). TorsoMap realigned to the v4 node schema (tissue/container/identity/guid/is_a; the stale `tiers`/`root` refs would have crashed on select). Turntable frames re-rendered crisp (jpg -> png). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TzqvDqbFRzyx17EkLKBoZF
Operator: "connect to a triangle filled surface ... kurvenlineal over triangles
(Quadro/AutoCAD)" + "stay with highest quality, do NOT coarsen." The BodyParts3D
OBJs are already triangulated with per-vertex normals, so render the FACES filled
(THREE.Mesh) instead of sampling centroids as splat points. Vertex-cluster decimation
averages the normals per cell (the smooth "curve-ruler" over triangles); filled +
Phong it is a solid surface (ivory bone, red muscle, blue cartilage) — the Open 3D
Man material aesthetic, no sequins, no fog. Decisively beats the surfel splat.
- bake_torso_mesh.py: SPM1 indexed mesh (vert 21 B [pos|normal|rgb|opacity|node_row]
+ tri 12 B), reuses the is_a tissue classifier. Highest quality: cell 3.6 mm =
1.40M triangles / 594K verts (29 MB), NOT coarsened.
- TorsoMesh.tsx: live THREE.Mesh, custom Phong shader (smooth interpolated normals,
two-sided via gl_FrontFacing), opaque depth-tested. uFloor cuts the skin/flesh
shell so the anatomy shows; S toggles skin.
- Routes: /torso-live = the filled mesh hero; /torso-splat = the opaque surfels
(kept for comparison); /torso turntable frames re-rendered at 1.4M (jpg->png).
Note: torso.mesh is a 29 MB binary asset (quality-first, per operator). Web-weight
coarsening was rejected ("skull looks terrible" at cell 5.5 mm).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TzqvDqbFRzyx17EkLKBoZF
Records the triangle-surface answer (SPM1 filled mesh beats surfels) and the read-only ndarray investigation: simd_soa MultiLaneColumn + Morton cascade pattern + cesium SSE/HLOD + aabb.rs batch cull are all gaussian-free and reusable for an opaque mesh; the splat3d render half (project/tile/raster/sh) is gaussian+alpha-blend. Verdict PARTIAL — strong reusable substrate, must build the triangle render layer on top (no gaussian coupling). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TzqvDqbFRzyx17EkLKBoZF
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
|
Caution Review failedThe pull request is closed. ℹ️ Recent review info⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: ⛔ Files ignored due to path filters (40)
📒 Files selected for processing (19)
📝 WalkthroughWalkthroughUpgrades the torso anatomical visualization pipeline from SPL2 gaussian soft-splats to SPL3 opaque triangle surfels and adds a new SPM1 filled triangle mesh view. The bake tooling switches from ChangesTorso v4 is_a-primary pipeline
Sequence Diagram(s)sequenceDiagram
participant Browser
participant TorsoMesh
participant decodeSpm1
participant ThreeJS as Three.js Scene
Browser->>TorsoMesh: fetch /torso.mesh
Browser->>TorsoMesh: fetch /torso.mesh.manifest.json
TorsoMesh->>decodeSpm1: ArrayBuffer (SPM1 binary)
decodeSpm1-->>TorsoMesh: Mesh {positions, normals, colors, opacity, index}
TorsoMesh->>ThreeJS: mount(container, mesh) — build BufferGeometry + ShaderMaterial
ThreeJS-->>TorsoMesh: fps via onStats (adaptive pixel ratio loop)
Browser->>TorsoMesh: keydown "S"
TorsoMesh->>ThreeJS: update uFloor uniform (skin shell toggle)
sequenceDiagram
participant OBJ_Files as BodyParts3D OBJ Files
participant load_isa
participant read_obj_tris
participant knn_radii
participant encode_SPL3 as SPL3 Encoder
participant Outputs as torso.splat + nodes.json + manifest.json
load_isa-->>encode_SPL3: is_a ownership, tissue_of, isa_dn, tissue_color
OBJ_Files->>read_obj_tris: parse triangle surfels
read_obj_tris->>knn_radii: per-surfel positions → per-surfel disk radii
knn_radii->>encode_SPL3: quantized scale, scale_max (99th pct)
encode_SPL3->>Outputs: SPL3 binary + enriched JSON sidecar files
Estimated code review effort🎯 4 (Complex) | ⏱️ ~60 minutes Possibly related PRs
Poem
✨ Finishing Touches📝 Generate docstrings
Warning There were issues while running some tools. Please review the errors and either fix the tool's configuration or disable the tool if it's a critical failure. 🔧 ESLint
ESLint install failed due to a network error. Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
AdaWorldAPI
pushed a commit
that referenced
this pull request
Jun 24, 2026
…er toggles + transparency Adds MY interactive cockpit viewer, additive to the other session's /torso* (#57/#58 — their TorsoMesh/torso.mesh are untouched; main.tsx gains only one import + one route). - cockpit_bake (new fma bin): bakes the full body → cockpit/public/fma_body.mesh, the same SPM1 wire the cockpit already decodes, but the per-vertex `opacity` byte carries a clean LAYER id (1 skin·2 muscle·3 organ·4 skeleton·5 vessel·6 nerve·7 connective·8 other) instead of a continuous alpha — so each layer can be toggled exactly. Color is the converged `tissue` byte (is_a); geometry is the BodyParts3D is_a OBJ set, vertex-cluster decimated (a Rust port of bake_torso_mesh.py's cluster_decimate). 329K verts, 745K tris. - cockpit/src/FmaBody.tsx (new): loads /fma_body.mesh, renders the solid THREE.Mesh with the same (-x,z,y) upright decode as TorsoMesh (head-up), and a shader that gates each layer via uEnabled[9] (indexed by the vertex layer id) + a global alpha. UI: a button per layer (skin/muscle/organ/skeleton/vessel/nerve) toggling visibility, and a solid↔transparent switch. Skin off by default so the anatomy shows. - cockpit/src/main.tsx: + route /fma-body (additive; /torso* unchanged). - README: fix the v3 image (the removed dot-graph PNG → the skeleton render) and document cockpit_bake + /fma-body. Verified: cockpit `npm run build` (tsc && vite build) is green (102 modules). The live browser render was not exercised in this environment (no browser) — the component mirrors the working TorsoMesh exactly and shares its decode, so it should render head-up on deploy. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GJ4NVBSjq1w5h7RmTbVafb
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Real human anatomy (BodyParts3D, FMA-keyed) rendered in the q2 cockpit, ending on a filled, smooth, Gouraud/Phong triangle surface — solid ivory bone, red muscle, blue cartilage, the Open 3D Man material aesthetic. Three render generations, each correcting the last:
Cockpit views
/torso-live— the hero: liveTHREE.Mesh, custom two-sided Phong shader (gl_FrontFacing), opaque depth-tested. Skin/flesh shell gated off so the anatomy shows; press S to toggle skin./torso— turntable of the same surface (pre-rendered frames)./torso-splat— the opaque surfel point cloud (kept for comparison)./torso-map— the splat as the GUID substrate: click a structure → its FMA node → tissue /container:identityGUID / is_a DistinguishedName / partonomy. (Also fixes a latent crash: it read staletiers/rootfields the v4 bake no longer emits.)Bake tooling (
crates/osint-bake/tools/)bake_torso_mesh.py— SPM1 indexed triangle mesh (vert 21 B[pos|normal|rgb|opacity|node_row]+ tri 12 B), vertex-cluster decimation with cell-averaged normals (smooth), is_a tissue colour.bake_torso_splat.py— SPL3 per-triangle surfels with k-NN local-spacing disk sizing (replaced a hand-rolledsqrt(stride)over-inflation; the surfel path is retained).export_isa_classification.py— full FJ→tissue/system/GUID/DN table (1658 concepts / 2234 meshes), one classifier shared with the gaussian-render session.Substrate alignment (investigated read-only)
The operator asked whether the Morton tile pyramid +
simd_soa.rs"gridlake" can give batched-shader synergy for the mesh without the gaussian artifacts. Verdict: PARTIAL —simd_soa.rsMultiLaneColumn, the Morton min/max cascade pattern, cesiumsse/hlod, andaabb.rsbatch cull are all gaussian-free and reusable; the splat3d render half (project/tile/raster/sh) is gaussian + alpha-blend, so an opaque triangle render needs its own tiler + projection + z-buffer raster on top. This matches ndarray's own 3DGS render-depth-certification plan, which already frames the mesh as the exact geometry anchor ("CAD mesh says wall is here") and the 3DGS gaussian as the radiance skin — unified by the geometry-agnostic HHTL / 4×4-SoA / depth-certificate substrate. The "opaque now (mesh), gaussian later (3DGS skin)" split is that planned reality-skinned-CAD/BIM architecture. Full write-up:claude-notes/research/2026-06-24-splat-on-substrate-jc-clam-cesium-hhtl.md.Notes
cockpit/public/torso.meshis a 29 MB binary (quality-first, per operator direction).🤖 Generated with Claude Code
https://claude.ai/code/session_01TzqvDqbFRzyx17EkLKBoZF
Generated by Claude Code
Summary by CodeRabbit
New Features
Bug Fixes